Intelligent feature elimination via Monte Carlo approach for machine learning-based groundwater nitrate prediction and health risk assessment

• This study presents a Monte Carlo Feature Selection (MCFS) approach for predicting nitrate concentrations in groundwater. • Five machine learning models were employed to predict nitrate concentrations. • A total of 147 drinking water samples from 12 different locations were analyzed. • Non-carcinogenic risk associated with NO₃ groundwater was assessed. This study presents a Monte Carlo Feature Selection (MCFS) approach for predicting nitrate concentrations in groundwater, utilizing predefined randomized subsets of five features and applying the ExtraTrees regression algorithm. Feature selection is governed by stringent criteria, including an R² performance threshold and feature occurrence frequency. The approach enhanced prediction accuracy by 8% and achieved a 50% reduction in the number of input features. Among the variables, bicarbonate (HCO₃⁻) was identified as the most influential, appearing in 66.4% of the high-performing feature subsets. An analysis of 147 groundwater samples collected from 12 sites indicated considerable health risks for infants and children, as evidenced by HQ values exceeding 1(HQ > 1). Among the five evaluated models of AdaBoost, Random Forest, Extra Trees, Decision Tree, and CatBoost, the Extra Trees algorithm exhibited the highest predictive performance, maintaining an R² of approximately 0.98 both prior to and following feature selection, thereby validating the robustness of the employed method. The optimized model preserved only the most consistently selected features, achieving a high level of predictive accuracy.